Abstract
Bees are the main pollinators of the world and are dying at an alarming rate. Being able to classify them and study their habits is of paramount importance. Crowdsourced datasets are preferred methods for gathering data about the current state of bee populations in their natural environment. Such images, however, may be problematic to use due to large volume of images that place strain on the experts' capabilities of identifying the species. We propose a method to identify regions of interest in an image containing a bee and to correctly classify the species of the bee. In addition, the procedure works on large crowdsourced datasets (we worked with BeeSpotter) with minimal manual annotation and data augmentation. Our approach is capable of addressing two genus and related bee species and records 91% correct classification. A limitation of the BeeSpotter dataset is labeling just one bee per image which may contain two or more bees. We overcome this issue by classifying all bees even in cases of two genus. Finally, the proposed approach is compared with two other recent works which report similar accuracy, but are limited with stricter image preprocessing or photographic setup. (C) 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (https://crativecommons.org/licenses/by-nc-nd/4.0) Peer-review under responsibility of the scientific committee of KES International.